Integration of the as-built reality represented using point cloud and as-designed generally represented using Building Information Modeling is an integral process in progress monitoring and more general Digital Twin (DT) activities. An up-to-date as-built model helps save time, and resources and allows for informed decision-making. Alignment helps update the old or outdated as-designed model to the current standards. But due to lack of digitization and BIM not being readily available, comparing point clouds taken at different phases during the building lifecycle provides an alternative approach for progress monitoring and compliance checking. Traditionally, this process involves overlaying recent point cloud data with the designed model or previous point cloud data, identifying common elements, and analyzing the differences. However, this method has limitations, especially when elements are far apart or have different shapes. This study utilized a graph-based learning method to link semantic instances between two point clouds of a scene. Graph-based representations were derived and enriched, and multi-modal encoders were leveraged to link two semantically same building elements. Subsequently, this method was employed to investigate PCD-BIM integration. The trained multi-modal architecture showed promising results with an Mean Reciprocal Rank (MRR) score of 0.99 when normal scenes were analyzed and an MRR score of 0.97 when deviated scenes were fed to the model. The base case had an MRR score of 0.79, showing promising results of graph-enriching methods, an increase of 25%, and demonstrating that the method is robust to positional deviations. Using PCD-PCD to generalize PCD-BIM did not provide sufficient insights but highlighted the need for its training module. The results showcase the proposed solution’s potential for integrating PCD-PCD in the built environment. The best-performing model finds links between the same building elements without processing individual points, significantly reducing the computing time. This leads to the downstream task of finding node correspondences and registration being computationally cheaper and faster. Keywords: Point Cloud, BIM, Entity Alignment, Graph Attention Network (GAT), Multi-Modal Encoder, Semantic, MRR
«
Integration of the as-built reality represented using point cloud and as-designed generally represented using Building Information Modeling is an integral process in progress monitoring and more general Digital Twin (DT) activities. An up-to-date as-built model helps save time, and resources and allows for informed decision-making. Alignment helps update the old or outdated as-designed model to the current standards. But due to lack of digitization and BIM not being readily available, comparing...
»